follow test14, 1. 粗层求解加速(多 block 点积归约, coarse_solve 533->42ms, 10x); 2. 大最粗层 fused 回退; 3. dev76 测试套件(verify PASS, sweep 14/16>=1.5); [opencode].
follow test14, 1. 粗层求解加速(多 block 点积归约, coarse_solve 533->42ms, 10x); 2. 大最粗层 fused 回退; 3. dev76 测试套件(verify PASS, sweep 14/16>=1.5); [opencode].